Poor Deployment Can Waste 99% of AI’s Potential

The better your AI models, the more value you lose in deployment

In Case You Missed It

Can AI Deploy Itself? Deploying a technology has always cost far more than buying it. AI may be the first that can help with its own deployment.

TL;DR

• AI investments convert to bottom-line value via a six-link Deployment Chain: Reimagine the work, Reallocate the resources, Reshape the organisation, Rewire the systems, Realise the value, Repeat the cycle.

• The gap between AI’s potential and delivered value can be explained by weak performance along the Deployment Chain. For example, only 21% of AI adopters have redesigned workflows, and only around 6% report material earnings impact. Weakness multiplies across the chain.

• Using survey data to estimate the losses at each stage of the Deployment Chain illustrates that up to 99% of AI’s potential value is being lost to leaky deployment.

• Better models without better deployment increase the value left on the table. Model capability is continuously improving while the Deployment Chain improves only when it is rebuilt. The best response is to fix the weakest link, not to upgrade the model.

In 1712, at a coalworks near Dudley, Thomas Newcomen’s engine began pumping water out of a flooded coal mine. Some 600 more had been put to work in Britain by 1775. But the engine wasted almost everything it burned. Some of the fire’s heat went straight up the chimney. More was lost turning water into steam. The cylinder was cooled and reheated on every stroke, the biggest single loss. The beam and the pump rods wasted a little more. In the end, less than 1% of the energy in the coal lifted any water. For half a century this was tolerated because the engines sat at coal mines and burned cheap coal.

AI is today’s frontier technology, as steam was in the eighteenth century. In common with early steam engines, little of AI’s full potential is finding its way to the bottom line. Between buying the model and banking its value sits the Deployment Chain: Reimagine the work. Reallocate the resources. Reshape the organisation. Rewire the systems. Realise the value. Repeat the cycle.

Reimagine the work

Reimagine is the first link in the deployment chain. AI enters existing workflows that are designed around people, not machines. Those workflows often have many variants, where colleagues have found ingenious workarounds to friction points. During the Reimagine stage, end-to-end workflows need to be reconceived around a new division of labour, where AI and human both have clear territory. Tried and tested tools such as Lean and Six Sigma underperform as the requirement is for a fundamentally new workflow, not an optimised version of the old one. McKinsey finds only 21% of AI adopters have redesigned any workflow. Call the pass rate from this link in the chain 20–35%.

Reallocate the resources

Reallocate commits capital and talent and specifies where in the P&L the benefit lands. For example, an approved efficiency business case should lead to someone’s budget being cut. NatWest is explicit about where benefits accrue. Its 2025 AI, cloud and simplification work freed roughly £100 million of investment capacity, and rather than let that saving dissipate it recycled the capacity back into the transformation. Few get this far: KPMG finds only around 7% of firms have established a clear view of AI’s return, and 42% have only partial sight of what they are even spending. Call the pass rate 25–45%.

Reshape the organisation

Reshape redesigns roles, skills, decision rights and accountabilities: who decides, who handles exceptions, who is accountable when the model is wrong. Reshape and Rewire (the subsequent link in the chain) are co-dependent. The organisation specifies what the system must support, and the live system reveals what the organisation must change. Job design alone is not sufficient to drive adoption. People must be retrained, and must learn when to accept, challenge and override. Deloitte finds only about a third have redesigned roles or career paths around AI, with retraining, not role redesign, the most common response. Call the pass rate 30–50%.

Rewire the systems

Rewire embeds that division of labour in production systems and data. Demos run on curated inputs, but deployments meet systems of record, half-clean data and undocumented exceptions. S&P Global found 42% of enterprises abandoning most of their AI initiatives in 2025, up from 17% a year earlier. Nearly half of proofs of concept are scrapped before production. Call the pass rate 15–55%, depending on what counts as production.

Realise the value

Realise is the conversion of operational improvements into bottom-line outcomes: cost removed, capacity redeployed, customer value uplifted. This is where the commitments made during Reallocate fall due. In Denmark, 25,000 workers across 7,000 workplaces adopted chatbots widely, saved about 3% of their time, and showed no measurable change in earnings or hours. The high-performing organisations in McKinsey’s sample, attributing material earnings impact to AI, are about 6% of adopters. Call the pass rate 5–20%.

Repeat the cycle

Repeat reruns the five gates as models improve and the economics shift. Stanford’s AI Index puts the fall in inference cost at more than 280-fold in under two years. As models improve, all links in the chain need to be upgraded to realise full value. For example, Intercom’s Fin customer service bot resolved a quarter of conversations at its 2023 launch and averages 67% today, roughly a percentage point improvement a month. Only 27% of enterprises have the MLOps tooling to run a Repeat loop at all. AI deployment should be an operating rhythm, not a programme with an end date.

The price of a weak link

We can define the Deployment Yield as the share of an AI investment’s potential value that reaches the P&L. Because value must flow through every link in the chain, total yield from an investment can be estimated by multiplying the individual yields. Five gates yielding 90% lead to an end-to-end yield of about 60%, whilst five gates at 80% yield about one third. The arithmetic is illustrative, but it explains a common pattern in technology deployments, whereby every team appears to perform well but the net impact is underwhelming.

Hence, with survey proxies, we can estimate the Deployment Yield for AI investments. Using the optimistic end of each range, we do not even reach 1%. The true yield will be higher, because the links are correlated: organisations disciplined enough to redesign the work tend also to commit the benefit and wire the systems. But not much higher. McKinsey finds about 6% of adopters attributing material earnings impact to generative AI, and BCG puts its ‘future-built’ share at 5%.

The same arithmetic suggests how yield can be increased. Improving any one link lifts the whole chain’s yield in proportion to that link’s weakness. Ten percentage points added to a 40% link raise total yield by 25%. The same ten points at an 80% link raise it by 12.5%. Effort pays most at the weakest link. So based on our estimates, the first places to look for improvement are the Realise and Reimagine links.

There are two fair objections. The first is that the estimates are flaky: they are self-reported consultancy surveys, and the links are not independent. Both points are true, but even with more optimistic assumptions, the yield remains in single digits, in line with the outcome surveys. The second objection is that every general-purpose technology experiences a ‘lag’ effect in its first years. This has been well proven by others. What’s different this time is the pace of change. Factories were rebuilt around the dynamo for over forty years while the technology itself changed slowly. With AI, we have a technology that is evolving faster than we are learning how to deploy it. Models are improving exponentially, whilst deployment improves only when someone rebuilds a link in the chain. Self-deploying agents may eventually rebuild links in the chain themselves; for now, that work is done by people, at organisational speed. Strong models do not compensate for weak links. In fact, as models improve, the value left on the table increases, unless the Deployment Chain is rebuilt.

The Playbook for Better Deployment

• Plan and cost the whole chain. Build every AI business case across all six Rs, ensuring the whole chain is fully costed.

• Estimate where the value is leaking. Look for workflow unchanged, benefit uncommitted, roles unaltered, exceptions unhandled, value unbanked.

• Fix the weakest link first. In a multiplicative chain the highest return sits at the weakest link. Resist the instinct to upgrade models until you’ve fixed the deployment chain.

• Give the chain one owner. The links cross functional lines: technology owns Rewire, HR owns Reshape, finance owns Realise – so by default nobody owns the whole.

• Run the chain as a rhythm, not a programme. Fund the Repeat laps as an ongoing expenditure, so that Deployment Yield is progressively increased and value unlocked.

Questions the Board should ask

• What is our Deployment Yield and where is it leaking?

• Do our business cases price the whole Deployment Chain, or only the model?

• Who is accountable for improving deployment?

The League Table

Steam-engine efficiency improved fastest once engineers started measuring it. From 1811 a monthly journal, Lean’s Engine Reporter, published how much water every Cornish mine engine lifted for each bushel of coal it burned. Once the numbers were public, the engineers competed, and by 1840 the best engine was four times more efficient. Two centuries on, the lesson holds. The best way to improve outcomes is not to buy a more powerful engine. It is to mend the weakest link.

Sources & Notes

Several figures below are directional rather than definitive. Historical duty figures depend on the coal bushel used (84 lb in Smeaton’s survey, about 94 lb in the later Cornish series) and are quoted on each source’s own basis; survey findings are largely self-reported; NatWest figures are the bank’s own account; the model-cost estimate comes from a venture investor; the chain arithmetic is the author’s illustration; and the per-link pass rates are the author’s planning ranges, built from self-reported surveys whose definitions differ.

• Pass-rate ranges. The per-link pass rates (20–35, 25–45, 30–50, 15–55 and 5–20%) and the end-to-end multiplication are an author synthesis for planning purposes, not observed benchmarks. Definitions differ across the underlying surveys, most of which are self-reported consultancy data, and the links are positively correlated, so naive multiplication understates the true yield. The observed single-digit shares of material-impact adopters are the empirical check.

• The Newcomen engine. First commercial installation 1712 at the Conygree coalworks near Dudley; around 600 engines at work in Britain by 1775 (John Kanefsky and John Robey, “Steam Engines in 18th-Century Britain: A Quantitative Assessment”, Technology and Culture, 1980). Steam was condensed inside the working cylinder, chilling and reheating the metal every stroke, and overall thermal efficiency was around 0.5%, so well under 1% of the coal’s energy reached the water (Britannica; standard histories). The engines were economic mainly at collieries burning near-free pit-head coal; Cornwall, with no coalfield, imported coal at high cost, which is why the county led demand for efficiency.

• Workflow redesign. McKinsey & Company, The State of AI: How Organizations Are Rewiring to Capture Value, March 2025 (surveyed July 2024): only about 21% of adopters had fundamentally redesigned at least some workflows. The November 2025 edition, The State of AI in 2025: Agents, Innovation, and Transformation (1,993 respondents), finds workflow redesign the organisational change most strongly associated with reported EBIT impact, with high performers nearly three times as likely as others to have redesigned workflows.

• Task length. METR, 2025: reports rapid growth in the length of tasks frontier models can complete reliably.

• Role redesign. Deloitte, State of AI in the Enterprise, 2026 edition: 3,235 leaders across 24 countries, fielded August to September 2025. About a third of organisations report redesigning roles or career paths around AI, with education the most common talent response; a narrower question on redesigning jobs themselves finds only 16% have done so (84% have not). A consultancy survey; findings are self-reported.

• Abandonment. S&P Global Market Intelligence, Voice of the Enterprise: AI & Machine Learning, Use Cases 2025 (451 Research survey of 1,006 IT and business leaders across North America and Europe, fielded late 2024): 42% of companies abandoned most of their AI initiatives, up from 17% a year earlier, with an average 46% of proofs of concept scrapped before production; the top cited obstacles were cost, data privacy and security. The same survey puts MLOps tool adoption at 27%, up from 24%, the capability proxy used for the Repeat link.

• Pilot-to-production. IDC’s AI CIO Playbook with Lenovo, 2025: 88% of observed proofs of concept did not reach wide-scale deployment, roughly 4 in 33 graduating. Vendor-commissioned with IDC fieldwork. The spread against S&P’s 46% scrappage figure is definitional: ‘wide-scale deployment’ is a higher bar than ‘any production’.

• The Danish null. Anders Humlum and Emilie Vestergaard, “Large Language Models, Small Labor Market Effects”, NBER working paper, May 2025, revised October 2025. Across roughly 25,000 workers and 7,000 workplaces, average time savings were near 3%, with no measurable effects on earnings or hours two years after adoption.

• Single digits. McKinsey & Company, The State of AI in 2025: Agents, Innovation, and Transformation (November 2025, 1,993 respondents): about 6% of adopters qualify as high performers attributing 5% or more of EBIT to generative AI, and the report describes enterprise-wide bottom-line impact as rare. BCG, The Widening AI Value Gap (September 2025, 1,250 executives): 5% of companies rated ‘future-built’ and generating substantial value. Both are consultancy surveys with a commercial interest in AI adoption.

• The dynamo. Paul A. David, “The Dynamo and the Computer: An Historical Perspective on the Modern Productivity Paradox”, American Economic Review Papers and Proceedings, May 1990. Electric motors passed half of US factory drive only in the early 1920s, four decades after commercial electricity, as factories rebuilt around unit drive.

• Model economics. Guido Appenzeller, Andreessen Horowitz, November 2024, estimates inference cost for constant capability falling roughly 10× per year. The estimate comes from an investor with an interest in the build-out it implies; the direction is widely accepted.

• Inference deflation. Stanford HAI, AI Index Report 2025: the cost of querying a model at GPT-3.5-level performance (64.8 on MMLU) fell from $20.00 per million tokens in November 2022 to $0.07 by October 2024, a more than 280-fold reduction in under two years; depending on task, LLM inference prices have fallen 9–900× per year.

• Fin. Intercom reports its Fin agent resolving conversations in the mid-20s at its March 2023 launch, rising to a 67% average across roughly 7,000 customers by December 2025, about a percentage point a month (Archana Agrawal, Intercom president, GTM Now podcast, February 2026; Intercom engineering account, 2026). The 25-to-above-80% spread across customers, and its dependence on ongoing content and escalation-rule maintenance, is from Intercom’s own optimisation guidance. Intercom sells Fin and charges per resolution, so the figures are the vendor’s own; one competitor analysis puts typical production rates at 45–53%.

• NatWest. NatWest Group reporting and CIO commentary, February 2026: £1.2 billion invested in technology, data and AI in 2025, with £100 million of investment capacity freed. These are the bank’s own figures.

• KPMG. Global AI Pulse survey, Q2 2026: more than 2,000 senior leaders across 20 countries at firms with revenues above US$50 million; 7% report having established ROI from AI, and 42% report only partial visibility into their AI spending. A consultancy survey from a firm that sells AI services; figures are self-reported.

• Lean’s Engine Reporter. From 1811, Joel Lean and later his family published monthly duty reports for Cornish pumping engines, engine by engine with engineers named, strokes taken from counters kept in locked boxes; publication continued until 1904. Reported duty rose from about 20 million (best 22.3 million at Wheal Alfred, 1811) to about 90 million by 1840. Alessandro Nuvolari (“Collective invention during the British Industrial Revolution: the case of the Cornish pumping engine”, Cambridge Journal of Economics, 2004) reads the episode as collective invention: open comparative measurement, after Watt’s patent expired in 1800, drove the improvement the monopoly years had not.